Peng Cheng 的 AI 实验室Peng Cheng's AI Lab

让智能体
真正做事
Agents that
actually work

我着迷于大模型与智能体。这里记录我的实验与作品:多智能体编排、工具调用、RAG,以及让 AI 读文档、写代码、看懂设备。 I'm fascinated by LLMs and agents. This is where I keep my experiments — multi-agent orchestration, tool use, RAG, and AI that reads documents, writes code and understands machines.

01 — 作品Work

用 AI 构建的系统Things I've built with AI

01多智能体 · 代码生成Multi-agent · Codegen

DeepBMS

让多个智能体读懂需求文档,自动写出可编译的电池管理系统代码。 Agents that read requirement documents and write compilable battery-management firmware.

  • LLM-first 解析 Excel / Word 需求文档为结构化参数
  • 模板化 C 代码注入、状态机适配与 CAN 协议生成
  • AutoGen 分阶段编排,每一步产出都可审查
  • LLM-first parsing of Excel / Word specs into structured parameters
  • Template C injection, state-machine adaptation, CAN code generation
  • Staged AutoGen orchestration with reviewable outputs
PythonAutoGenLLMCCAN
查看源码View source →
02诊断智能体 · RAGDiagnostic agent · RAG

储能运维智能体Operations Agent

一个会看数据、会查资料、会做决定的运维智能体:告警 → 诊断 → 检索 → 处置。 An agent that reads telemetry, looks things up and acts: alarm → diagnose → retrieve → respond.

  • 三级故障规则诊断,联动禁放与继电器保护动作
  • 知识库问答、人工确认工单、诊断报告一键导出
  • 健康评分、多租户与 Docker 私有化部署
  • Level-3 fault diagnosis wired to discharge lock and relay protection
  • Knowledge-base QA, human-confirmed tickets, exportable reports
  • Health scoring, multi-tenancy and Docker on-prem deployment
FastAPIVue 3RAGPostgreSQLDocker
查看源码View source →
03LLM · 文档生成LLM · Documents

SOP-IGS

让大模型参与文档生产:照片归类、AI 校验与 Word 自动排版。 LLMs in document production: photo grouping, AI review and automated Word layout.

  • 基于固定 Word 母版,按项目裁剪测试章节
  • 消防、BMS、温升等照片分组上传与自动排版
  • Word COM Worker 生成可编辑 DOCX,LLM 只建议不阻断
  • Tailors test chapters from a fixed Word master
  • Grouped photo upload and auto-layout for fire, BMS, thermal tests
  • Word COM worker outputs editable DOCX; the LLM advises, never blocks
PythonFastAPIWord COMPostgreSQL
查看源码View source →
04多模型 · 市场分析Multi-LLM · Markets

StockAgent

多个大模型协作解读市场,再用因子与回测去验证。 Several LLMs read the market together, then factors and backtests check them.

  • 多 LLM 生成个股分析报告
  • 17+ 选股因子与向量化回测,计入 T+1、费用与涨跌停
  • 微服务 + gRPC 架构,Tushare 行情定时同步
  • Multi-LLM stock analysis reports
  • 17+ factors and a vectorised backtester with T+1, fees and price limits
  • Microservices over gRPC with scheduled Tushare sync
FastAPILangChaingRPCVue 3MongoDB
查看源码View source →

02 — 关于About

关于 AI,
我相信的几件事
What I believe
about AI

我是 Peng Cheng。比起和模型聊天,我更好奇它能替人把多少事真正做完:读懂一份几十页的协议、写出能编译的代码、在凌晨三点判断一次告警。这个网站是我的实验场,记录做过的东西,也记录踩过的坑。

I'm Peng Cheng. More than chatting with models, I'm curious how much real work they can finish on their own — reading a forty-page protocol, writing code that compiles, triaging an alarm at 3 a.m. This site is my lab: what I've built, and what I've learned the hard way.

01

智能体不是聊天框An agent is not a chat box

真正有用的 AI 要能调用工具、读写系统,并为结果负责。对话只是入口。Useful AI calls tools, reads and writes real systems, and owns its results. Conversation is just the front door.

02

规则与模型各司其职Rules and models, each in their place

确定的交给代码,不确定的交给模型,关键决定留给人。这样的系统才可靠。Certainty goes to code, ambiguity to the model, and the critical call to a human. That is what makes it dependable.

03

AI 正在走进物理世界AI is reaching the physical world

从文档、代码到电池与设备,模型的边界在不断外扩。我想站在这条边界上做东西。From documents and code to batteries and machines, the frontier keeps moving outward. That edge is where I like to build.

03 — 联系Contact

聊聊 AI 吧Let's talk AI

对智能体、大模型,或者某个有点疯狂的想法感兴趣?欢迎来信。Into agents, LLMs, or a slightly crazy idea? Drop me a line.